Machine Learning Approaches for Breast Cancer Survivability Prediction

Breast cancer is one of the leading causes of cancer death in women. If not diagnosed early, the 5-year survival rate of patients is just about 26\%. Furthermore, patients with similar phenotypes can respond differently to the same therapies, which means the therapies might not work well for some of them. Identifying biomarkers that can help predict a cancer class with high accuracy is at the heart of breast cancer studies because they are targets of the treatments and drug development. Genomics data have been shown to carry useful information for breast cancer diagnosis and prognosis, as well as uncovering the disease’s mechanism. Machine learning methods are powerful tools to find such information. Feature selection methods are often utilized in supervised learning and unsupervised learning tasks to deal with data containing a large number of features in which only a small portion of them are useful to the classification task. On the other hand, analyzing only one type of data, without reference to the existing knowledge about the disease and the therapies, might mislead the findings. Effective data integration approaches are necessary to uncover this complex disease. In this thesis, we apply and develop machine learning methods to identify meaningful biomarkers for breast cancer survivability prediction after a certain treatment. They include applying feature selection methods on gene-expression data to derived gene-signatures, where the initial genes are collected concerning the mechanism of some drugs used breast cancer therapies. We also propose a new feature selection method, named PAFS, and apply it to discover accurate biomarkers. In addition, it has been increasingly supported that, sub-network biomarkers are more robust and accurate than gene biomarkers. We proposed two network-based approaches to identify sub-network biomarkers for breast cancer survivability prediction after a treatment. They integrate gene-expression data with protein-protein interactions during the optimal sub-network searching process and use cancer-related genes and pathways to prioritize the extracted sub-networks. The sub-network search space is usually huge and many proteins interact with thousands of other proteins. Thus, we apply some heuristics to avoid generating and evaluating redundant sub-networks